In the verification of C programs by deductive approaches based on automated provers, some heuristics of separation analysis are proposed to handle the most difficult problems. Unfortunately, these heuristics are not sufficient when applied on industrial C programs: some valid verification conditions cannot be automatically discharged by any automated prover mainly due to their size and a high number of irrelevant hypotheses. This work presents a strategy to reduce program verification conditions by selecting their relevant hypotheses. The relevance of a hypothesis is the combination of separated static dependency analyzes based on graph constructions and traversals. The approach is applied on a benchmark issued from industrial program verification.